The useful thing about a bottleneck is its honesty. It tells you where the system has stopped pretending. In software, the celebrated arrival of generative AI made one task astonishingly cheap: producing more code. Then the bill slid a few seats down the table. More code meant more pull requests, more tests, more security checks, more continuous-integration failures, and more senior engineers trying to decide whether a plausible change was also a safe one.
Ali-Reza Adl-Tabatabai had seen versions of this movie long before anyone called the cast agentic. His career moved from optimizing compilers to virtual machines, from reliability organizations to the tools used by thousands of developers. Again and again, speed created a second problem: how to preserve understanding, correctness, and confidence once the machine moved faster.
Gitar, the company he started after leaving Uber, was his answer to the newest version. Its agents were built for what happens after code appears. They review changes, investigate failures, propose fixes, test those fixes, and work inside the pull-request machinery that engineering teams already use. The aim is not a bot that leaves an eloquent comment and wanders off. It is a workflow that advances.
“Generation produces code; validation makes it trustworthy.”Ali-Reza Adl-Tabatabai
Chapter one / Seeing through optimization
The debugger came before the agent
In 1996, Adl-Tabatabai completed a Ph.D. in computer science at Carnegie Mellon. His dissertation tackled an awkward bargain in computing. Compiler optimizations make programs run better, but they rearrange, remove, and transform operations in ways that can make source-level debugging confusing. A variable may be unavailable. A breakpoint may no longer correspond neatly to the execution the programmer imagined. The faster program can become the less legible program.
His work proposed techniques that let a debugger present the original source faithfully whenever possible and disclose the effects of optimization when it could not. He led development of cmcc, a retargetable optimizing C compiler with generators for MIPS, SPARC, and iWarp. He also worked on Omniware, a language-independent virtual machine for mobile code. There was even a campus-navigation wearable with a head-mounted display, speech input, and GPS, built by a graduate-student team he managed. It sounds like a prop sent backward from a more recent decade.
The subject was technical, but the instinct was broadly human: complexity should not force the user to surrender comprehension. Later chapters changed the machinery, not the instinct.
Ph.D. at Carnegie Mellon on debugging globally optimized code.
Led programming-systems research spanning compilers, runtimes, multicore systems, and hardware support.
Worked on Facebook's HipHop VM and its just-in-time compiler for PHP.
Engineering Director at Google, with Site Reliability Engineering responsibilities across products and cloud infrastructure.
Led Uber's Developer Platform organization.
Founded Gitar, raised $9 million, then joined Sonar through its acquisition of the company.
Chapter two / Scale changes the question
A platform is what everyone else stands on
At Intel Labs, Adl-Tabatabai directed the Programming Systems Lab and worked on technologies for transactional memory, parallel systems, compilers, and runtimes. By 2012, when he spoke about Facebook's HipHop project, he had published more than 40 papers and held 37 patents. Gitar's later launch materials put his patent count above 90. The bibliography is not incidental decoration. It shows a career spent at the border where software abstractions meet hardware consequences.
Facebook's HipHop work attacked PHP performance first by compiling the language into C++, then through a dedicated virtual machine and just-in-time compilation. Google brought a different kind of scale. As an engineering director, Adl-Tabatabai oversaw Site Reliability Engineering across communications, photos, social products, cloud systems, and technical infrastructure. Reliability stopped being a final checkpoint and became an organizational discipline.
Uber made the developer experience itself the system under examination. Adl-Tabatabai led the Developer Platform organization, responsible for frameworks and tools used daily across the company. In a 2023 keynote, the scale arrived in numbers large enough to become architectural constraints.
The same presentation described more than 100 million lines in a monorepo, seven programming languages, and a roughly 200-person platform team. At that scale, a mildly irritating tool becomes a company-wide tax. A slow feedback loop repeats thousands of times. A noisy check steals attention with industrial efficiency. Standardization, measurement, and fast feedback are less about tidiness than preserving the ability of engineers to think.
This is why Adl-Tabatabai talks about developer experience as a competitive advantage. Good tools keep developers in flow and automate the grunt work. Fragmented tooling does the opposite: it recruits engineers as part-time coordinators for systems that were supposedly bought to help them.
Chapter three / The queue moves downstream
AI made writing abundant. Confidence stayed scarce.
Adl-Tabatabai left Uber after six and a half years. In his launch post, he used a photograph from his final trip to the office, mentioned taking time off, then introduced Gitar and the former colleagues joining him. Gautam Korlam, a principal engineer on Uber's Developer Platform team, became his co-founder and technical counterpart. Their first formulation was broad: build a modern end-to-end development platform.
The market soon supplied a sharper problem. Coding assistants and autonomous agents could produce changes faster than teams could absorb them. Every change still had to travel through review, continuous integration, security checks, approvals, and merging. None of those obligations vanished because the first draft arrived quickly. The bottleneck had simply moved from implementation to confidence.
Gitar focused on validation instead of joining the crowded race to generate the next block of code. Its review agent could find an issue and prepare a fix. Its CI agent could analyze logs, group related failures, identify a likely cause, and attempt a repair on the branch. Platform teams could define organization-specific checks in natural language. The product's promise was active: do the work across tools, maintain context, and continue until the workflow had evidence of success.
“Developers want fast, noise-free tools that keep them in flow and automate the grunt work.”Ali-Reza Adl-Tabatabai
That emphasis on low noise matters. Code review is already an argument about limited attention. An AI system that produces generic warnings can make the queue longer while congratulating itself for participation. Adl-Tabatabai's standard is closer to execution: Can the agent handle multiple steps? Can it use tools, preserve context, validate its output, and move the task forward without requiring a person to supervise every gesture?
Chapter four / Earning the merge button
Trust climbs. It does not teleport.
The revealing part of his plan is not full autonomy. It is the route he proposes for getting there. Adl-Tabatabai describes four levels of trust, each based on performance at the level before it. Organizations do not begin by handing an agent the keys to the main branch. They begin by watching whether it finds real problems without flooding the team with false alarms.
At the first level, the agent detects. At the second, it remediates and gets the checks to pass. At the third, it approves changes that satisfy policies defined by the team. At the fourth, it merges, including the delicate work of resolving conflicts as commit volume increases. Human involvement narrows toward decisions with risk, setting intent, reviewing outcomes, and handling exceptions.
This is ambitious automation presented with a researcher's suspicion of unearned certainty. Break the work into smaller steps. Set boundaries. Validate continuously. Keep people where judgment carries consequence. The agent earns a wider remit through evidence, not vocabulary.
Chapter five / A fast second act
Five weeks from arrival to acquisition
On April 15, 2026, Gitar emerged from stealth with $9 million led by Venrock and joined by Sierra Ventures. It already named customers including SoFi, Cadence, Revyl, XFactor.io, and Sphinx. The funding was intended for engineering, product, enterprise sales, and developer awareness. Adl-Tabatabai said success over the following year would mean enterprise adoption, recognition among developers, and agents assuming more of the review, remediation, and merge workflow.
Five weeks later, Sonar acquired the company. The fit was almost diagrammatic. Sonar brought deterministic analysis across syntax, data flows, logic, architecture, and dependencies. Gitar brought AI-native review, fixes, and CI iteration. Adl-Tabatabai and Korlam joined Sonar to continue leading Gitar, which remained available as a standalone product while also moving into Sonar's broader verification platform.
For a founder, acquisition can make a tidy ending. Here it reads more like a change in scale. Gitar's mission acquired access to Sonar's installed base, and Sonar acquired a team obsessed with the messy work between a generated change and a trustworthy merge. At the AI Engineer World's Fair that June, Adl-Tabatabai's scheduled subject said it plainly: “The Verification Gap Nobody Talked About.”
His career has never really left that gap. A debugger trying to explain optimized code, a lab making parallel systems programmable, a virtual machine making PHP fast, a reliability organization keeping products upright, a platform team smoothing the path of thousands of developers: each chapter lives between capability and confidence.
AI has made the capability spectacularly visible. Adl-Tabatabai is working on the less theatrical half of the bargain. Software still has to survive contact with tests, policies, dependencies, colleagues, and production. The future of coding may contain far more agents. His wager is that the useful ones will know how to check their work.